{ "id": 395256, "title": "Are we allowed to use fragments from other papyri?", "url": "/competitions/vesuvius-challenge-ink-detection/discussion/395256", "author_name": "", "post_date": "2023-03-16T15:11:31.384120300Z", "votes": 4, "comment_count": 7, "views": 0, "content": "

e.g. from here https://www.lib.umich.edu/locations-and-hours/papyrology-collection/find-and-access-papyri

\n

In the rules it says no additional fragments of the papyri we’re working with (I guess to prevent Data leakage from train to test sets?) but I wondered if we can use other papyri.

", "messages": [ { "id": "2184713", "postDate": "03/16/2023 15:11:31", "content": "

e.g. from here https://www.lib.umich.edu/locations-and-hours/papyrology-collection/find-and-access-papyri

\n

In the rules it says no additional fragments of the papyri we’re working with (I guess to prevent Data leakage from train to test sets?) but I wondered if we can use other papyri.

", "rawMarkdown": "e.g. from here https://www.lib.umich.edu/locations-and-hours/papyrology-collection/find-and-access-papyri\n\nIn the rules it says no additional fragments of the papyri we’re working with (I guess to prevent Data leakage from train to test sets?) but I wondered if we can use other papyri.", "votes": null }, { "id": "2184756", "postDate": "03/16/2023 15:48:59", "content": "

That is a good question. We don't allow using any data from other Herculaneum papyri, indeed to prevent data leakage, but if you somehow want to use non-Herculaneum papyri for something, then that is allowed.

\n

I'm curious, what are you thinking of using those for? As far as I'm aware there aren't high-resolution 3D X-ray volumes available for most papyri, only photos?

", "rawMarkdown": "That is a good question. We don't allow using any data from other Herculaneum papyri, indeed to prevent data leakage, but if you somehow want to use non-Herculaneum papyri for something, then that is allowed.\n\nI'm curious, what are you thinking of using those for? As far as I'm aware there aren't high-resolution 3D X-ray volumes available for most papyri, only photos?", "votes": null }, { "id": "2184871", "postDate": "03/16/2023 16:47:08", "content": "

I want to mention that there are publicly available CT datasets for various other manuscripts (scrolls, books, etc.). We don't immediately see how some of these might be useful (different materials, different resolutions, etc), but you, our clever contestants, might think otherwise! I'll see about getting a list of public manuscript datasets in our docs.

\n

As an example, the supporting data* for the original ink-id paper includes a proxy papyrus scroll with carbon ink applied at various thicknesses. Though the resolution is lower than the Herculaneum data in this contest, it might be useful as training data. Note that the PHercParis2 dataset in that archive is off limits though, since it's another Herculaneum fragment.

\n

*I just created my account here, so I can't post a link yet. DOI: 10.17605/OSF.IO/ZDKN4

", "rawMarkdown": "I want to mention that there are publicly available CT datasets for various other manuscripts (scrolls, books, etc.). We don't immediately see how some of these might be useful (different materials, different resolutions, etc), but you, our clever contestants, might think otherwise! I'll see about getting a list of public manuscript datasets in our docs. \n\nAs an example, the supporting data* for the original ink-id paper includes a proxy papyrus scroll with carbon ink applied at various thicknesses. Though the resolution is lower than the Herculaneum data in this contest, it **might** be useful as training data. Note that the PHercParis2 dataset in that archive is off limits though, since it's another Herculaneum fragment. \n\n*I just created my account here, so I can't post a link yet. DOI: 10.17605/OSF.IO/ZDKN4", "votes": null }, { "id": "2189307", "postDate": "03/20/2023 11:50:38", "content": "

Just to further clarify: We are not even allowed to use pictures from previous attempts to unroll the papyri of Herculaneum? (e.g. https://en.wikipedia.org/wiki/Herculaneum_papyri#/media/File:Tesoro_letterario_di_Ercolano_p27.jpg or https://en.wikipedia.org/wiki/Herculaneum_papyri#/media/File:Tesoro_letterario_di_Ercolano_p29.jpg)

", "rawMarkdown": "Just to further clarify: We are not even allowed to use pictures from previous attempts to unroll the papyri of Herculaneum? (e.g. https://en.wikipedia.org/wiki/Herculaneum_papyri#/media/File:Tesoro_letterario_di_Ercolano_p27.jpg or https://en.wikipedia.org/wiki/Herculaneum_papyri#/media/File:Tesoro_letterario_di_Ercolano_p29.jpg)", "votes": null }, { "id": "2191070", "postDate": "03/21/2023 17:14:57", "content": "

As per the current rules, that is indeed not allowed. I'm curious what you want to use these photos for? How are they useful without corresponding X-ray data?

\n

If there are particular Herculaneum photos that are useful for some reason, we could consider loosening the rules a little. But there would have to be a strong argument, since it would make the rules more complicated.

", "rawMarkdown": "As per the current rules, that is indeed not allowed. I'm curious what you want to use these photos for? How are they useful without corresponding X-ray data?\n\nIf there are particular Herculaneum photos that are useful for some reason, we could consider loosening the rules a little. But there would have to be a strong argument, since it would make the rules more complicated.", "votes": null }, { "id": "2191090", "postDate": "03/21/2023 17:30:11", "content": "

I would use such images (without X-Ray counterparts) in a GAN style approach to generate ink masks which do contain greek letters. However you have to be very careful not to create hallucinations. If this would mean to much effort in adopting the rules, I could also use images of greek letters from other papyri.

\n

And sorry to bother you with another question: Can the additional data (e.g. scans of complete scrolls) from https://scrollprize.org/data be used in this competition, or is scrollprize.org considered as external to the Kaggle competition.

\n

Thank you in advance

", "rawMarkdown": "I would use such images (without X-Ray counterparts) in a GAN style approach to generate ink masks which do contain greek letters. However you have to be very careful not to create hallucinations. If this would mean to much effort in adopting the rules, I could also use images of greek letters from other papyri.\n\nAnd sorry to bother you with another question: Can the additional data (e.g. scans of complete scrolls) from https://scrollprize.org/data be used in this competition, or is scrollprize.org considered as external to the Kaggle competition.\n\nThank you in advance", "votes": null }, { "id": "2191122", "postDate": "03/21/2023 18:06:04", "content": "

There's a lot of photos of papyri out there, so if you can use others then that would be better.

\n

You're allowed to use the additional data from https://scrollprize.org/data in the Kaggle competition.

", "rawMarkdown": "There's a lot of photos of papyri out there, so if you can use others then that would be better.\n\nYou're allowed to use the additional data from https://scrollprize.org/data in the Kaggle competition.", "votes": null }, { "id": "2192056", "postDate": "03/22/2023 11:58:14", "content": "

Yes, the solution above seems a profitable way forward for sure. Using other greek texts to predict the most likely character/word/phrase from what can be visually determined, and then apply that back as a pixel mask seems to be the best way.

\n

Hallucinations are probable, but if the results and confidence metrics are explainable, it seems the technique would be useful.

\n

For example, GPT4 tells me that

\n
\n

In Ancient Greek, one of the most common 2-grams is \"το\" (to), which means \"the\" in English. Other common 2-grams include \"ὁ δε\" (ho de), meaning \"but he\", \"εις το\" (eis to), meaning \"into the\", \"του δε\" (tou de), meaning \"but his\", and \"και ὁ\" (kai ho), meaning \"and the\".

\n
\n

If one was able to discern some scattering of pixels that hinted at these 2 grams, you might be able to roughly fill in the rest, at least probabilistically. No idea if this is accurate, you'd need to confirm what the most probable 2-grams for ancient Herculaneum papyri might be, but it's useful as an example.

\n

I was also reading on wiki that https://en.wikipedia.org/wiki/Mount_Vesuvius

\n
\n

These papyri contain a large number of Greek philosophical texts. Large parts of Books XIV, XV, XXV, and XXVIII of the magnum opus of Epicurus, On Nature and works by early followers of Epicurus are also represented among the papyri.[19] Of the rolls, 44[citation needed] have been identified as the work of Philodemus of Gadara, an Epicurean philosopher and poet. The manuscript \"PHerc.Paris.2\" contains part of Philodemus' On Vices and Virtues.[2]

\n
\n

Again, asking GPT4 (again, purely an example, confirm for accuracy):

\n
\n

However, based on the surviving fragments, the most common characters used in the copies of \"On Nature\" are likely to be similar to the most common characters used in other Ancient Greek texts. This means that the letter \"α\" (alpha) is still expected to be one of the most common characters, followed by \"ε\" (epsilon) and \"ο\" (omicron).

\n
\n

Even if these aren't philosophical texts it might be reasonable to assume that some of the characters and text could be similar to them. If they are something else entirely of course, this way might not work very well and only produce hallucinations.

\n

The temptation might be to restrict the comp to pixel identification without the assistance of context, but I'd argue that could be short sighted. Solutions which work in tandem could be much more effective than simply trying to do pixel detection alone

\n

Maybe even it would be counter-productive to focus on pixel detection alone at this stage. Imagine for example you were able to use a contextual solution to provide better training/validation data, which would lead to solutions for determining such things as inked drawings that might exist in the scrolls.

", "rawMarkdown": "Yes, the solution above seems a profitable way forward for sure. Using other greek texts to predict the most likely character/word/phrase from what can be visually determined, and then apply that back as a pixel mask seems to be the best way.\n\nHallucinations are probable, but if the results and confidence metrics are explainable, it seems the technique would be useful.\n\nFor example, GPT4 tells me that \n\n> In Ancient Greek, one of the most common 2-grams is \"το\" (to), which means \"the\" in English. Other common 2-grams include \"ὁ δε\" (ho de), meaning \"but he\", \"εις το\" (eis to), meaning \"into the\", \"του δε\" (tou de), meaning \"but his\", and \"και ὁ\" (kai ho), meaning \"and the\".\n\nIf one was able to discern some scattering of pixels that hinted at these 2 grams, you might be able to roughly fill in the rest, at least probabilistically. No idea if this is accurate, you'd need to confirm what the most probable 2-grams for ancient Herculaneum papyri might be, but it's useful as an example.\n\nI was also reading on wiki that https://en.wikipedia.org/wiki/Mount_Vesuvius\n\n>These papyri contain a large number of Greek philosophical texts. Large parts of Books XIV, XV, XXV, and XXVIII of the magnum opus of Epicurus, On Nature and works by early followers of Epicurus are also represented among the papyri.[19] Of the rolls, 44[citation needed] have been identified as the work of Philodemus of Gadara, an Epicurean philosopher and poet. The manuscript \"PHerc.Paris.2\" contains part of Philodemus' On Vices and Virtues.[2]\n\nAgain, asking GPT4 (again, purely an example, confirm for accuracy):\n\n>However, based on the surviving fragments, the most common characters used in the copies of \"On Nature\" are likely to be similar to the most common characters used in other Ancient Greek texts. This means that the letter \"α\" (alpha) is still expected to be one of the most common characters, followed by \"ε\" (epsilon) and \"ο\" (omicron).\n\nEven if these aren't philosophical texts it might be reasonable to assume that some of the characters and text could be similar to them. If they are something else entirely of course, this way might not work very well and only produce hallucinations.\n \nThe temptation might be to restrict the comp to pixel identification without the assistance of context, but I'd argue that could be short sighted. Solutions which work in tandem could be much more effective than simply trying to do pixel detection alone\n\nMaybe even it would be counter-productive to focus on pixel detection alone at this stage. Imagine for example you were able to use a contextual solution to provide better training/validation data, which would lead to solutions for determining such things as inked drawings that might exist in the scrolls.", "votes": null } ], "comments": [ { "id": 2184756, "author_name": "jpposma", "author_url": "", "post_date": "03/16/2023 15:48:59", "content": "

That is a good question. We don't allow using any data from other Herculaneum papyri, indeed to prevent data leakage, but if you somehow want to use non-Herculaneum papyri for something, then that is allowed.

\n

I'm curious, what are you thinking of using those for? As far as I'm aware there aren't high-resolution 3D X-ray volumes available for most papyri, only photos?

", "votes": null, "replies": [ { "id": 2189307, "author_name": "patrickaiforfun", "author_url": "", "post_date": "03/20/2023 11:50:38", "content": "

Just to further clarify: We are not even allowed to use pictures from previous attempts to unroll the papyri of Herculaneum? (e.g. https://en.wikipedia.org/wiki/Herculaneum_papyri#/media/File:Tesoro_letterario_di_Ercolano_p27.jpg or https://en.wikipedia.org/wiki/Herculaneum_papyri#/media/File:Tesoro_letterario_di_Ercolano_p29.jpg)

", "votes": null, "replies": [ { "id": 2191070, "author_name": "jpposma", "author_url": "", "post_date": "03/21/2023 17:14:57", "content": "

As per the current rules, that is indeed not allowed. I'm curious what you want to use these photos for? How are they useful without corresponding X-ray data?

\n

If there are particular Herculaneum photos that are useful for some reason, we could consider loosening the rules a little. But there would have to be a strong argument, since it would make the rules more complicated.

", "votes": null, "replies": [ { "id": 2191090, "author_name": "patrickaiforfun", "author_url": "", "post_date": "03/21/2023 17:30:11", "content": "

I would use such images (without X-Ray counterparts) in a GAN style approach to generate ink masks which do contain greek letters. However you have to be very careful not to create hallucinations. If this would mean to much effort in adopting the rules, I could also use images of greek letters from other papyri.

\n

And sorry to bother you with another question: Can the additional data (e.g. scans of complete scrolls) from https://scrollprize.org/data be used in this competition, or is scrollprize.org considered as external to the Kaggle competition.

\n

Thank you in advance

", "votes": null, "replies": [ { "id": 2191122, "author_name": "jpposma", "author_url": "", "post_date": "03/21/2023 18:06:04", "content": "

There's a lot of photos of papyri out there, so if you can use others then that would be better.

\n

You're allowed to use the additional data from https://scrollprize.org/data in the Kaggle competition.

", "votes": null, "replies": [] }, { "id": 2192056, "author_name": "kaggleqrdl", "author_url": "", "post_date": "03/22/2023 11:58:14", "content": "

Yes, the solution above seems a profitable way forward for sure. Using other greek texts to predict the most likely character/word/phrase from what can be visually determined, and then apply that back as a pixel mask seems to be the best way.

\n

Hallucinations are probable, but if the results and confidence metrics are explainable, it seems the technique would be useful.

\n

For example, GPT4 tells me that

\n
\n

In Ancient Greek, one of the most common 2-grams is \"το\" (to), which means \"the\" in English. Other common 2-grams include \"ὁ δε\" (ho de), meaning \"but he\", \"εις το\" (eis to), meaning \"into the\", \"του δε\" (tou de), meaning \"but his\", and \"και ὁ\" (kai ho), meaning \"and the\".

\n
\n

If one was able to discern some scattering of pixels that hinted at these 2 grams, you might be able to roughly fill in the rest, at least probabilistically. No idea if this is accurate, you'd need to confirm what the most probable 2-grams for ancient Herculaneum papyri might be, but it's useful as an example.

\n

I was also reading on wiki that https://en.wikipedia.org/wiki/Mount_Vesuvius

\n
\n

These papyri contain a large number of Greek philosophical texts. Large parts of Books XIV, XV, XXV, and XXVIII of the magnum opus of Epicurus, On Nature and works by early followers of Epicurus are also represented among the papyri.[19] Of the rolls, 44[citation needed] have been identified as the work of Philodemus of Gadara, an Epicurean philosopher and poet. The manuscript \"PHerc.Paris.2\" contains part of Philodemus' On Vices and Virtues.[2]

\n
\n

Again, asking GPT4 (again, purely an example, confirm for accuracy):

\n
\n

However, based on the surviving fragments, the most common characters used in the copies of \"On Nature\" are likely to be similar to the most common characters used in other Ancient Greek texts. This means that the letter \"α\" (alpha) is still expected to be one of the most common characters, followed by \"ε\" (epsilon) and \"ο\" (omicron).

\n
\n

Even if these aren't philosophical texts it might be reasonable to assume that some of the characters and text could be similar to them. If they are something else entirely of course, this way might not work very well and only produce hallucinations.

\n

The temptation might be to restrict the comp to pixel identification without the assistance of context, but I'd argue that could be short sighted. Solutions which work in tandem could be much more effective than simply trying to do pixel detection alone

\n

Maybe even it would be counter-productive to focus on pixel detection alone at this stage. Imagine for example you were able to use a contextual solution to provide better training/validation data, which would lead to solutions for determining such things as inked drawings that might exist in the scrolls.

", "votes": null, "replies": [] } ] } ] } ] } ] }, { "id": 2184871, "author_name": "csparker", "author_url": "", "post_date": "03/16/2023 16:47:08", "content": "

I want to mention that there are publicly available CT datasets for various other manuscripts (scrolls, books, etc.). We don't immediately see how some of these might be useful (different materials, different resolutions, etc), but you, our clever contestants, might think otherwise! I'll see about getting a list of public manuscript datasets in our docs.

\n

As an example, the supporting data* for the original ink-id paper includes a proxy papyrus scroll with carbon ink applied at various thicknesses. Though the resolution is lower than the Herculaneum data in this contest, it might be useful as training data. Note that the PHercParis2 dataset in that archive is off limits though, since it's another Herculaneum fragment.

\n

*I just created my account here, so I can't post a link yet. DOI: 10.17605/OSF.IO/ZDKN4

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2184713": "e.g. from here https://www.lib.umich.edu/locations-and-hours/papyrology-collection/find-and-access-papyri\n\nIn the rules it says no additional fragments of the papyri we’re working with (I guess to prevent Data leakage from train to test sets?) but I wondered if we can use other papyri.", "2184756": "That is a good question. We don't allow using any data from other Herculaneum papyri, indeed to prevent data leakage, but if you somehow want to use non-Herculaneum papyri for something, then that is allowed.\n\nI'm curious, what are you thinking of using those for? As far as I'm aware there aren't high-resolution 3D X-ray volumes available for most papyri, only photos?", "2184871": "I want to mention that there are publicly available CT datasets for various other manuscripts (scrolls, books, etc.). We don't immediately see how some of these might be useful (different materials, different resolutions, etc), but you, our clever contestants, might think otherwise! I'll see about getting a list of public manuscript datasets in our docs. \n\nAs an example, the supporting data* for the original ink-id paper includes a proxy papyrus scroll with carbon ink applied at various thicknesses. Though the resolution is lower than the Herculaneum data in this contest, it **might** be useful as training data. Note that the PHercParis2 dataset in that archive is off limits though, since it's another Herculaneum fragment. \n\n*I just created my account here, so I can't post a link yet. DOI: 10.17605/OSF.IO/ZDKN4", "2189307": "Just to further clarify: We are not even allowed to use pictures from previous attempts to unroll the papyri of Herculaneum? (e.g. https://en.wikipedia.org/wiki/Herculaneum_papyri#/media/File:Tesoro_letterario_di_Ercolano_p27.jpg or https://en.wikipedia.org/wiki/Herculaneum_papyri#/media/File:Tesoro_letterario_di_Ercolano_p29.jpg)", "2191070": "As per the current rules, that is indeed not allowed. I'm curious what you want to use these photos for? How are they useful without corresponding X-ray data?\n\nIf there are particular Herculaneum photos that are useful for some reason, we could consider loosening the rules a little. But there would have to be a strong argument, since it would make the rules more complicated.", "2191090": "I would use such images (without X-Ray counterparts) in a GAN style approach to generate ink masks which do contain greek letters. However you have to be very careful not to create hallucinations. If this would mean to much effort in adopting the rules, I could also use images of greek letters from other papyri.\n\nAnd sorry to bother you with another question: Can the additional data (e.g. scans of complete scrolls) from https://scrollprize.org/data be used in this competition, or is scrollprize.org considered as external to the Kaggle competition.\n\nThank you in advance", "2191122": "There's a lot of photos of papyri out there, so if you can use others then that would be better.\n\nYou're allowed to use the additional data from https://scrollprize.org/data in the Kaggle competition.", "2192056": "Yes, the solution above seems a profitable way forward for sure. Using other greek texts to predict the most likely character/word/phrase from what can be visually determined, and then apply that back as a pixel mask seems to be the best way.\n\nHallucinations are probable, but if the results and confidence metrics are explainable, it seems the technique would be useful.\n\nFor example, GPT4 tells me that \n\n> In Ancient Greek, one of the most common 2-grams is \"το\" (to), which means \"the\" in English. Other common 2-grams include \"ὁ δε\" (ho de), meaning \"but he\", \"εις το\" (eis to), meaning \"into the\", \"του δε\" (tou de), meaning \"but his\", and \"και ὁ\" (kai ho), meaning \"and the\".\n\nIf one was able to discern some scattering of pixels that hinted at these 2 grams, you might be able to roughly fill in the rest, at least probabilistically. No idea if this is accurate, you'd need to confirm what the most probable 2-grams for ancient Herculaneum papyri might be, but it's useful as an example.\n\nI was also reading on wiki that https://en.wikipedia.org/wiki/Mount_Vesuvius\n\n>These papyri contain a large number of Greek philosophical texts. Large parts of Books XIV, XV, XXV, and XXVIII of the magnum opus of Epicurus, On Nature and works by early followers of Epicurus are also represented among the papyri.[19] Of the rolls, 44[citation needed] have been identified as the work of Philodemus of Gadara, an Epicurean philosopher and poet. The manuscript \"PHerc.Paris.2\" contains part of Philodemus' On Vices and Virtues.[2]\n\nAgain, asking GPT4 (again, purely an example, confirm for accuracy):\n\n>However, based on the surviving fragments, the most common characters used in the copies of \"On Nature\" are likely to be similar to the most common characters used in other Ancient Greek texts. This means that the letter \"α\" (alpha) is still expected to be one of the most common characters, followed by \"ε\" (epsilon) and \"ο\" (omicron).\n\nEven if these aren't philosophical texts it might be reasonable to assume that some of the characters and text could be similar to them. If they are something else entirely of course, this way might not work very well and only produce hallucinations.\n \nThe temptation might be to restrict the comp to pixel identification without the assistance of context, but I'd argue that could be short sighted. Solutions which work in tandem could be much more effective than simply trying to do pixel detection alone\n\nMaybe even it would be counter-productive to focus on pixel detection alone at this stage. Imagine for example you were able to use a contextual solution to provide better training/validation data, which would lead to solutions for determining such things as inked drawings that might exist in the scrolls." }, "source": "meta" }